How Typewise got into YC after pivoting to B2B productivity

Typewise, a Swiss firm, demonstrates the value of perseverance: The team behind text prediction technology has received support from Y Combinator and will be in the cohort pitching to investors during the accelerator’s summer 2022 demo day early next month. The team’s fascination with improving typing productivity began as a consumer keyboard-focused side business more than five years ago.

After changing its direction to completely concentrate on the B2B market, Typewise was awarded a position in YC (along with its customary $500,000 financing). According to co-founder David Eberle, this was done to meet the need for typing efficiency benefits in fields like customer support and sales.

He tells TechCrunch, “Last year we identified where this made the most sense.” “Consumers use WhatsApp to write a few phrases here and there, so being 20% or 30% quicker or having a few less errors doesn’t really matter to them. However, in organisations, particularly those that involve a lot of writing, such as customer service and sales, even single digit percentages count a lot and double digit percentages much more.

Quality matters a lot since it is customer-facing communication and because it might affect a brand’s reputation, the author continues. “So, ultimately, it got us into YC.”

Typewise received $1 million in what it termed as a seed in 2020, but Eberle says the company now views it as more of a pre-seed and plans to seek further funding when it pitches investors in September.

Despite concentrating entirely on B2B, the team’s work on Typewise’s consumer app—which has received more than 2 million downloads—was not in vain. According to Eberle, it assisted them in “fine-tuning” their AI models, which allowed them to submit a second patent application earlier this year for technology that can anticipate full sentences as opposed to only the words that would come after them.

Sentence prediction is now a key selling factor, driving efficiency improvements that, in the case of one early Typewise client, a package delivery/logistics firm with whom it has been working the longest, exceeded 35% (on average) a few weeks after the company began using the technology.

Other early users come from a variety of sectors, such as e-commerce, retail, and insurance.

Customers can use Typewise’s technology as a browser extension, which, according to Eberle, connects to a server-side API where the AI is housed. However, the entire system is intended to run on top of customer CRM platforms like Salesforce or Zendesk, integrating Typewise’s text predictions into pertinent client applications like email or live chat (i.e., places where business agents are talking, by text, to their own customers).

According to Eberle, the 10 or so early adopters of its MVP, which released this spring, are seeing increases of 10% to 20% on average as a result of incorporating the text prediction technology into their workflow. However, he asserts that he is convinced that the higher number (35%) will be the norm rather than the exception when Typewise adjusts the model’s parameters or otherwise fine-tunes it in response to client data and demand (and as customer staff get accustomed to using the AI-powered text prediction tool).

When asked how Typewise differs from other text prediction technologies, Eberle notes that Typewise not only offers a base language model (it supports 40 languages, though early customers are primarily focused on English and German) but also retrains and fine-tunes its model using actual customer data. This implies that it can provide personalised predictions, which he claims are around 2.5 times more accurate than generic next word prediction AI found in mobile operating systems or email clients since it has been educated on user-specific data.

For instance, he explains, “we would look at all the support requests from the previous year or two and take them, and there’s a difficult filtering process (because sometimes you have to weed out poor quality language that you do not want to put into your training sets)).” If you compare our prediction to, say, a Gmail prediction, where the words are relatively typical, we get genuine material. The AI then iterates on the client data.

Due to the fact that, for instance, the language used in email communications between a company and its clients may vary significantly from that used in live text chat, Typewise may additionally partition its AI models based on the linguistic context (which is probably more fluid and informal, etc). In order to provide more contextually relevant (and hence fruitful) text predictions, it is thus organising a lot of consumer data inputs and datasets in the background, utilising machine learning technologies to assist it automate the required data structure.

Eberle reiterates, “It’s real content because we focus on a very specific use case,” and claims that this strategy provides it a distinct advantage over firms using generative language models, such as GPT-2 or GPT-3, to enable text prediction for their own B2B play.

He further emphasises that instead than forcing customers to upload vast amounts of consumer data, the solution has been designed such that the AI training process takes place inside the client’s systems. (Note: Analytics of the model’s performance may still need data to be provided back to Typewise, although Eberle claims it supports a few tiers so this process may not necessarily require real client content to be submitted.)

“There are certainly now a tonne of new businesses working on language aid, paraphrasing tools, attempting to make the language better, offering you advice [etc.], and many of them employ GPT-3 as their technology. They lack proprietary technologies. A [big telecom] or insurance firm, for instance, is not going to give up all of their customer conversations for you to train the AI, which is a drawback. Accordingly, we can practically deploy an instance of the AI into the customer’s IT infrastructure, and that way all the customer data stays with the enterprise but our AI kind of becomes a part of their data structure, he explains, adding that this is how we get around any IT security and data privacy issues that would probably otherwise make this pretty much impossible.

Given that Typewise’s text predictions must be able to update in real time during live text conversations in order to be effective (and not unpleasant) for the human agents the technology is endowing with superhuman typing speed skills. Latency is one of Typewise’s biggest challenges. Eberle claims that because latency optimization has been a priority, it has an advantage over text-generation technologies that have not placed a high priority on processing time reduction.

The use case, he says, “is that we’re communicating with a human person right now, and that’s quite different technologically from text production.” We cannot wait 300 or 500 milliseconds, which also appears like a very little amount of time, since our system requires an incredibly low latency. However, we must quickly update the prediction after each keystroke. Otherwise, it loses its human-useability. Therefore, the delay must be at least 50 milliseconds. As a result, one of the major limitations and difficulties in creating this is in the backdrop.

Could Typewise see further extending its technology to be able to completely automate customer-facing communications for its clients — at least in select areas, say like live chat for insurance sales or customer support emails for a package delivery company?

In response to this query, Eberle states that “something toward auto-reply” is one of the upcoming features on its roadmap, moving beyond the kinds of template-based, “pre-set” responses that can already cause an automated email with some degree of contextual relevance but where “the answer you get is always based on a pre-written template.”

That’s what their clientele don’t appreciate, according to what several businesses have told him, he adds. “How we view the future is that with more maturity… for a certain sort of ticket… ultimately we will see that for particular queries…,” Once the threshold of confidence is over a given amount, you may automate and declare that once the accuracy is above 99% or even more, you don’t need a human person anymore. The difference is that we would create such emails from scratch rather than using a text that has already been prepared. Word by word, we construct it. As a person would build it. That is how the AI works and how we created it.

Also Read: Law Commission proposes revolutionary rules for ownership of crypto tokens and NFTs

“Right now with this one customer that I mentioned, we got to 35% automation, which means that, on average, 35% of the emails were automatically created by Typewise. Hopefully, this number will rise. We are working on it,” he continues. “So right now it couldn’t yet finish a full email with five distinct content messages on its own without a human input,” the author said. “But clearly over time as those 35% move further up then that will be the case — and I believe that’s also the aim in the end.”

Of course, technologically advanced companies like Google and Microsoft are working on text prediction, but often just for their own products. Nevertheless, things may change. Eberle says, “So that’s what we’re monitoring attentively.”

Additionally, he foresees (ha!) that Grammerly may start to give text prediction. Although they don’t currently have text prediction, he says, “I’m fairly convinced that as the most useful language tool they will most certainly go into that field as well.” And I believe that personalization and the capability to do this, despite the issues with data privacy, is what truly sets us apart.

He also mentions the well-funded Wordtune product from AI21 Labs and the Dutch start-up Deep Desk as further competitors.

However, he also identifies “value add” aspects in Typewise’s pipeline that are expected to increase its distinction, such linking customer happiness ratings to language selections/styles in an effort to pinpoint the most effective linguistic strategies that result in contented clients.

Leave a Reply

Your email address will not be published. Required fields are marked *